Deedy Das

DeepSeek-OCR is the best OCR ever.

It parses this extremely hard to read handwritten letter written by mathematician Ramanujan in 1913 with a frightening degree of accuracy.

Not perfect, but beats former best dots ocr. Bonus points if you can spot the errors.

Try it here:

https://lnkd.in/gtuMpErT

We JUST solved a hard open math problem— the moving sofa problem!

“What is the area of the largest shape that can be manuevered through an L corner of width 1?“

On Dec 2, 2024, Korean Jineon Baek proved the answer is the 18 curve section Gerver's sofa with area 2.2195..!

Read the incredible proof here: https://lnkd.in/gUhbhtqZ

Meta is currently offering $2M+/yr in offers for AI talent and still losing them to OpenAI and Anthropic. Heard ~3 such cases this week.

The AI talent wars are absolutely ridiculous.

Today, Anthropic has the highest ~80% retention 2 years in and is the #1 (large) company top AI researchers wants to go.

Source: https://lnkd.in/gpaahyyR

Every single one of these $100M+ companies were started by alumni from a single Computer Science club in a non-American high school.

Cartesia
Inception Labs
General Catalyst CVF
Wispr Flow
Affinity
Snapdeal
Sugar
boAt

It's Exun Clan in Delhi Public School, RK Puram in India.

Using light as a neural network, as this viral video depicts, is actually closer than you think. In 5-10yrs, we could have matrix multiplications in constant time O(1) with 95% less energy. This is the next era of Moore's Law.

Let's talk about Silicon Photonics...

The core concept: Replace electrical signals with photons.

While current processors push electrons through metal pathways, photonic systems use light beams, operating at fundamentally higher speeds (electronic signals in copper are 3x slower) with minimal heat generation.

It's way faster.

While traditional chips operate at 3-5 GHz, photonic devices can achieve >100 GHz switching speeds. Current interconnects max out at ~100 Gb/s. Photonic links have demonstrated 2+ Tb/s on a single channel. A single optical path can carry 64+ signals.

It's way more energy efficient.

Current chip-to-chip communication costs ~1-10pJ/bit. Photonic interconnects demonstrate 0.01-0.1pJ/bit. For data centers processing exabytes, this 200x improvement means the difference between megawatt and kilowatt power requirements.


The AI acceleration potential is revolutionary.

Matrix operations, fundamental to deep learning, become near-instantaneous:
Traditional chips: O(n²) operations.
Photonic chips: O(1) - parallel processing through optical interference.

1000×1000 matmuls in picoseconds.

Where are we today?

Real products are shipping:
— Intel's 400G transceivers use silicon photonics.
— Ayar Labs demonstrates 2Tb/s chip-to-chip links with AMD EPYC processors. Performance scales with wavelength count, not just frequency like traditional electronics.

The manufacturing challenges are immense.

— Current yield is ~30%. Silicon's terrible at emitting light and bonding III-V materials to it lowers yield
— Temp control is a barrier. A 1°C change shifts frequencies by ~10GHz.
— Cost/device is $1000s

To reach mass production we need: 90%+ yield rates, sub-$100 per device costs, automated testing solutions, and reliable packaging techniques. Current packaging alone can cost more than the chip itself.

We're 5+ years from hitting these targets.

Companies to watch: ASML (manufacturing), Intel (data center), Lightmatter (AI), Ayar Labs (chip interconnects).

The technology requires major investment, but the potential returns are enormous as we hit traditional electronics' physical limits.

A tiny AI startup in Boston that sold for $500M in 2018 is the unlikely place 6 founders of near unicorns came from:

— OpenEvidence [$6B]: Daniel Nadler
— Surge [$25B]: Andrew Maubossin (CTO)
— Langchain [$1.25B]: Harrison Chase
— Suno [$1B+]: Mikey Shulman
— Thinking Machines [$12B]: Sam Shleifer (mts)
— Sequence Holdings: Neal Wu
— Chai Discovery: Joshua Meier
— at least 5 other startups (Genmo, Pebble, NewComputer), professors at MIT and more.

The company is called Kensho Technologies.

This is why you should join startups with smart people. The company might be average, but the outcomes can be spectacular.

Most teams of software engineers have 1 savant.

They just want to build. They have crazy ideas. They have lightning fast execution. And they don't like rules.

Your job as a manager is to protect your savant from management at all costs, steer them and let them fly.

At a small startup, everyone can be a savant and the company will be functional. As the company scales, you need to strike a balance.

There are brilliant engineers who aren't savants too. They're exceptional, but they thrive in structure, are very diligent. Not crazy though

If you don't let your savant fly and try to wedge them into conformity, they'll get bored and leave. You'll not only have lost a talent, you might permanently scar their brilliance.

One of the worst things about BigTech is that it kills the savant with rules. Don't do that.

HUGE Immigration News! $100k H-1B fee will not apply to students changing their status from F-1.

That mainly targets WITCH IT companies that do this, who are planning on using the L-1 instead.

If you studied in the US, the fee doesn’t affect you. I’m now positive on this rule!

Two 23 year old Indians just dropped the #2 open-weight AI voice model in the world, trained purely on free credits!

Maya1 is #20 globally, better than even Google's best. 3B params, runs on one GPU and does 20+ emotions with < 100ms latency

You can just do things.

[turn volume on to listen to audio on this post]
Source: https://lnkd.in/gk_iEaBF

Here's what Ilya saw.

The deposition of top exec in the Musk vs Altman lawsuit is the spiciest episode of this season of OpenAI. Huge attack on Sam's character and details about a Anthropic-OpenAI merger.

Linked the full transcript below:

Source: https://lnkd.in/gE3pwyXS

Meta just dropped this paper that spills the secret sauce of reinforcement learning (RL) on LLMs.

It lays out an RL recipe, uses 400,000 GPU hrs and posits a scaling law for performance with more compute in RL, like the classic pretraining scaling laws.

Must read for AI nerds.

Source: https://lnkd.in/gYZGnVBJ

The top 5 universities by number of unicorn founders in the US are:

#1 Stanford (135)
#2 MIT (100)
#3 Harvard (80)
#4 UC Berkeley (68)
#5 Indian Institute of Technology (65)

One college in India is #5 in the world for American unicorns. That's the power of immigration.

Note: You can roll up all the UCs in one for consistency and they would be #3, but the point remains

Source: Ilya Strebulaev

Every so often that one engineer who thinks “we should just rewrite this in Rust instead of using S3” is right and saves you $500k/yr..

..but now you need to staff a team of 5 engineers to maintain, monitor, upgrade and plan the release for this new service so 4yrs later the new CTO you hire from Amazon can come in and say “why the fuck did we build this when S3 exists?”

And the circle of life continues.

Fantastic engineering read though, seriously: https://lnkd.in/gJG6mvVH

This man who grew up in middle class India literally built the only startup outside US/China/Japan that sent rockets to space.

Here's the never-before-shared crazy story of Pawan Kumar Chandana:

> be middle class kid growing up in Vizag
> terrible student, get 51 marks in maths
> ambitious dad doesn't give up, puts you in IIT coaching
> fall in love with math and science
> go from worst to best in school
> crack IIT first attempt
> 2007: go to IIT KGP for mechanical engineering
> everyone around you chasing big packages, consulting, going abroad
> you just love rockets
> 2012: join ISRO as scientist straight out of campus
> pays peanuts
> loves it
> doing so well you want to retire here
> but entrepreneurial bug from IIT days never left
> dreams of building a global space company from India
> but no policy allowing private rockets, no funding environment for space
> 2018: quit anyway, ready to survive in rags
> googles "what is equity"
> zero connections
> cold DM Mukesh Bansal on LinkedIn
> he writes a $1.5M check
> COVID hits, seed capital running out
> Series A is a brutal struggle
> no fund invests
> founders of renewable company Greenko invests
> 2021: PM Modi opens up space sector to private players
> become first company to sign MoU with ISRO
> get largest check in Indian DeepTech, $51M
> Nov 18, 2022: launch Vikram-S: India's first suborbital private rocket
> reaches 90km
> Prime Minister Narendra Modi inaugurates your new facility
> grow to 1000 employees
> build India's largest private rocket factory at 200,000 sq ft
> valued at $527M
> first orbital launch, Vikram-1, planned in 2026
> will be one of ~5 companies globally regularly launching to orbit
> next: reusable spaceships
> Open Space for All

We talk a lot about Elon Musk and SpaceX, but imagine having NO prior wealth. Being in a developing country. Constantly saying no to money. Being unable to raise any venture funding multiple times. Being called crazy by everyone. And still having the perseverance to dream beyond the stars.

I'd call Pawan's story inspirational, but honestly it's more than that. It's beyond what I can fathom.

One of the most bizarre stories from Meta was when an intern brought the entire Facebook site down in 2013 just by using her initials as the internal username. This caused the entire production release to fail.

Her initials were “www”.

Total Anthropic Victory (for now). Everything they've done in 2026:
– $20B revenue run rate, about to cross OpenAI
– #1 on iOS and Android App Store
– 4x App downloads in last 10 days
– 2.6x web visits in 2mos
– >50% of enterprise AI subscriptions (in addition to LLM API spend)
– Anthropic > OpenAI on Google Trends
– Cowork causes SaaSpocalypse, tanks IT services
– Claude for Security tanks cyber stocks
– Claude COBOL modernizations tanks IBM
– Claude used in Maduro and Iran strikes.

And plus, Donald Knuth solves an open math problem with Claude Opus 4.6. If you're not using Claude, Cowork or Claude Code, well... everyone else is.

YouTube on its own would be a ~$500B company and #18 globally by market cap.

As of this quarter, it's a $60B ARR business (68% ads and 32% subscriptions) growing 15% YoY and accounts for 15% of Google revenue.

>2% of all human waking time is spent on YouTube.

One of the greatest products ever built.

Why we are not in an AI bubble in 4 charts:

— Multiples (P/E) are nowhere near .com level
— CapEx is growing but funded by cash flow
— Largest tech co valuations lower than 1999
— Concentration in the market isn't necessarily negative

Source: Coatue Oct 2025 State of AI Report

We're in the "what if Google does that" part of the AI cycle.

They can make
— cheaper models ($2/M in, $12/M out, just above GPT5.1, cheaper than Pro)
— better models (benchmarks)
— distribute products at no cost to billions of users (Gemini has >50% MAU of ChatGPT, Antigravity is free vs Cursor $20-40/mo)
— get good unit economics (own TPU, no reliance on Nvidia premiums), use it to retain premium talent cheaper

Of the BigTech giants, Amazon and Microsoft chose to be infra partners. Apple chose not to play. Meta shat the bed. Google is coming out on top.

Public reports inadvertently show that Anthropic is 7.8% owned by Amazon and up to 8.8% by Google.

— Amazon Q3 earnings says they gained $9.5B from Anthropic ownership which implies a 7.8% ownership stake.
— Google didn’t explicitly mention Anthropic but previous reports and “unrealized gains from non marketable equity” implies a 8.8% stake.

Claude Skills is a game changer for power users who are tired of prompting LLMs to do things exactly how they want.

Here's the 15 skills that Anthropic pre-packaged.

Each one is essentially a long pre-defined prompt for a specific task, a hybrid between a custom system prompt and a lightweight MCP.

So far, I've found it useful for basic tasks like:
— Consistent themes in my brand
— Generating text with my guidelines
— Using certain esoteric coding libraries the way they should be

I’ve heard 6 people tell me they’re doing this: Using vibe coding apps to algorithmically trade on the stock or crypto market.

They’re using 2-10x leverage to make up to +50% a month. This is a dangerous game to play.

Algo trading is the last thing I expected AI to democratize.

Source: https://lnkd.in/gjz2-TRG

Startups are fucking hard and here’s the math to prove it:

40% of startups die after a seed.
50% of the remainder die after a Series A.
60% of the remainder die after a Series B.
58% of the remainder die after a Series C.

Roughly ~2.5% after the seed are acquired, so not “dead”.

0.5-1% go IPO.

Assumptions: this is based on a ~10yr time horizon from 2016-2018 vintage. The exact numbers will change by vintage.

Credit: Peter Walker from Carta

🚨China's DeepSeek just dropped the only open-source model good enough at math to win IMO Gold, and a must-read report!

The key idea draws from things Karpathy and others have spoken about: move beyond “final answer RL” into a generator–verifier–meta-verifier loop in pure language.
– A verifier is RL-trained to score proofs.
– A meta-verifier checks the verifier’s critiques.
– A generator is RL-trained on verifier reward signals to write and self-check better proofs.

Because everything lives in natural language (no Lean), this recipe SHOULDextend to many verifiable domains: science, code, anywhere where checking is easier than solving!

Source: https://lnkd.in/gFytbq6T

Few people know that a small < 200 person Israeli startup Amazon bought 10yrs ago for $350M now powers the training of Anthropic’s Claude models, providing a cheaper alternative to Nvidia.

It’s called Annapurna Labs.

Source: https://lnkd.in/g-dmAgE4

This man dropped out of a no-name college in India to be a software engineer and by 33, worked his way up to being CEO of a $100M+ company in New York.

Here's the never-before-shared incredibly inspirational story of Ershad Kunnakkadan:

> be middle class kid in random state school in Kerala, India
> get really into computers
> senior hands you Ubuntu 8.04 CD, whole new world
> starts contributing to SMC (a malayalam computing group)
> gets into blogging cause SMC seniors are into it
> go to no-name small private college in Kerala
> spend more time in terminals than classrooms
> shell scripting contests, Linux admin, security, virus cleaning, bots, paper presentations
> doesnt see a point to college exams
> drops out after 2nd year, promises family "I will earn a degree somehow"
> lands internship at small software co
> grows into being an architect
> found security bugs in Github and Prezi
> reads "Hackers: Heroes of the Computer Revolution" and the "The Google Story", dreamt of being in the US one day
> does Google Summer of Code
> earns a degree remotely from Bharathiar University
> gets remote job at BigBinary
> moves back to Kochi to be near family
> involved in local free software circles and workshops, events, meetups
> building a quiet dense body of work over loud personal brand
> gets introduced to Gumroad as a consultant first
> joins Gumroad as a senior engineer
> gets married
> moves to Abu Dhabi to be closer to wife's family
> does the boring crucial stuff - scalability, security, payouts, infra
> grows to being a staff software engineer at Gumroad
> support millions of users and $1B+ in creator earnings
> just focusses on self-improvement
> never once thinks about promotion
> moves to New York City on an O-1 visa
> Gumroad looking for a new CEO
> board looks around and its clear who is best fit for the job
> become CEO of a $100M+ gmv business

I just love the story of Ershad. No brands, no pedigree, no MBA, no loudness. When I asked him the quality that got him here, he said "reliability". A truly kind, quiet and generous person. Who loves computers. Dropping out when you're rich is trendy in America, but to see someone Indian drop out and work their way up into the top role is pure inspiration.

Don't worry if you don't have all the accolades and ornaments you see in people who achieve your dreams. Be a good person, and be reliable.

Thanks @ershus for sharing your story!
Check out his blog here: https://lnkd.in/gwgvRATs

Both Cursor and Cognition (Windsurf) new models today are speculated to be built on Chinese base models!

– Cognition SWE-1.5 seems to be a customized (fine-tuned / RL) Zhipu’s GLM 4.6 model running on Cerebras.
– Cursor Composer has Chinese reasoning traces.

Interesting to see mature AI products move to owning the model!

Seedance 2.0 can go from manga to anime in one shot. This is the Moro arc of Dragon Ball Super from 2018-20.

Bytedance has officially passed the video Turing test.

We are about to see an explosion of demand for compute and supply of content like we have never seen before.

If you feel like giving up, you must read this never-before-shared story of the creator of PyTorch and ex-VP at Meta, Soumith Chintala

> from hyderabad public school, but bad at math
> goes to a "tier 2" college in India, VIT in Vellore
> rejected from all 12 universities for US masters despite 1420 on the GRE
> fuckit.jpg
> goes to the US anyway on a J-1 visa to CMU with no plan
> applies for masters (again) to 15 universities
> rejected from all except USC and with late admissions, NYU in 2010
> finds this guy called Yann LeCun (before he was famous)
> starts getting into open source
> rejected from all jobs including DeepMind
> only job is Amazon as test engineer
> his PhD mentor helps him get a job at a small startup (MuseAmi)
> rejected from DeepMind
> couldn't get H-1B because of J-1 home return issue; gets waiver through months of approval with USCIS and US State Dept
> very low on confidence
> In 2011/12 builds one of the fastest AI inference engines on phones
> rejected from DeepMind
> emailed Yann again and joins FAIR because of Torch7 open-source work
> scrapes through bootcamp at Facebook, struggling on an HBase task
> L8/L9 engineers at Facebook struggle to get ImageNet working
> figures out numerics / hyperparam issue as an L4
> first big win!
> FAIR goes well, runs 3 person torch7 team and co-creates PyTorch
> because of politics, management wants to shut down PyTorch
> cries-at-bar.jpg, literally
> eventually some people save PyTorch and it launches in 2017
> gets a EB-1 green card!
> the rest is history...

Think about that. He went to a tier 2 college. Was rejected from all Masters programs 2x. Rejected from every single job except Amazon test engineering. Rejected from DeepMind 3x. Nearly had his baby project shut down. Struggled with visa issues. After 12 years of failures (2005-17), he eventually rose to became a VP at Meta one of the most influential people in AI!

Soumith's story is one of resilience and he's living proof that no matter how down in the dumps you are, there's always hope.

Some lore. Here's this awesome presentation Soumith wrote about his work at MuseAmi in 2014 about getting convnets to run on phones: https://lnkd.in/g974MhKk

Meta just bought Manus for >$1B and it makes sense. ~8 Consumer AI apps hit $100M+ arr that aren’t big labs:

Perplexity: $20B
ElevenLabs: $6.6B
Lovable: $6.6B
Replit: $3B+
Suno: $2.5B
Gamma: $2.1B
Character: $1B+
Manus: $500M

Meta AI has ~no product. This was the cheapest, and most aligned.

The flippening is happening. Everything you read is more likely AI than not, like Reddit and amazon reviews.

New research measures what AI detectors actually work.

Pangram ranks #1 with < 0.5% false positive and negative rate. Works well with GPT-5, Grok and Sonnet 4.5 too.

This isn't sponsored. I've used it myself (you get 5 credits free) and it works pretty well. Just today, I asked GPT-5 to pull 10 quotes from research, and Pangram told me the intro was AI but each quote seemed human!

Source: https://lnkd.in/geKRWzFq

The billion dollar seed round era is upon us.

Lila Sciences
General Intuition
Periodic Labs
Thinking Machines
SSI
Sierra

A few software engineers at some of the best tech cos told me this week "My entire job these days is prompting Cursor or Claude Code with Opus 4.5 to do what I need and sanity checking it."

We've crossed some intangible threshold of AI generalizing to "most" software.

Source: https://lnkd.in/g8BFcgaM

I read every one of Anthropic’s job openings so you don’t have to. Turns out they’re working on way more than code.

Here are the 5 biggest new surprises (ABCDE):
— Audio: even though they’ve focused primarily on text, there’s a new role to work on “understanding and generating speech and audio” including speech language models and audio diffusion models
— Biology: accelerate progress in life sciences by 10x
— Cybersecurity: they have a data, RL and engg to make “AI powered products for cybersecurity”
— Discovery: build an AI Scientist that solves “scientific Artificial General Intelligence”
— Eyes / Vision: improve Claude’s vision and spatial capabilities

Google's new Gemini 3.0 Pro models, orionmist and lithiumflow, are in LMArena and they're so insanely good.

— Extremely good with web UI. No purple tinge. Very aesthetic, rounded corners.
— State of the art at SVGs and 3D minecraft meshes

Can't wait for these to launch!

There is huge title inflation happening in tech right now.

"Legacy" techcos are offering lofty titles to combat multi-million dollar offers from AI labs.

I counted > 500 "Head of" something at Stripe on LinkedIn, a ~10k person company: >5%. And they're far from alone.

BREAKING: OpenAI is planning a $1 TRILLION IPO by late 2026 / early 2027.

Today, I'm excited to launch my lifelong passion project, Grand Old Books!! 🚀

There are 1000s of beautiful novels of the past, not in English, locked up in old PDFs, with no physical copies left. We started with Indian texts and brought back 12 books in 6 languages with pictures and annotations.

This is, and will always be, completely free.

We can't let time wash away history.

Please comment to let me know what book you'd like to see added.

Please try it at: https://grandoldbooks.com/

Any feedback here is welcome!

IMO gold medalists are 50x more likely to win a Fields medal, the highest prize in Mathematics, than PhDs from top 10 schools.

The Power Law is everywhere for those with eyes to see.

Source: https://lnkd.in/g7pu6XTa

Elon's stance on the H-1B is very clear: "Some of the outsourcing companies have gamed the system on the H-1B front, and we should stop the gaming of the system..

but I'm NOT of the school of thought that we should shut down the H-1B program.. that would actually be very bad"

Inception Labs is the AI co everyone's sleeping on.

3 profs from Stanford, Cornell and UCLA just dropped Mercury 2, the first reasoning language (and code) diffusion model ever. It is 10x faster and the cheapest model for its quality.

They're not quite at frontier like the Claude 4.6 / 5.3s of the world but they are reshapinging pareto-frontier for price/quality and latency/quality.

If they get to frontier, they'd upend the current economics of large language models.

This Stanford professor just raised a $50M Seed and have built a 10x faster and 10x cheaper AI coding model with the performance of Gemini Flash / Haiku.

Inception’s Mercury model can implement games like Connect 4 from scratch in ~2s. The speed feels magical, like going from dial up to broadband. And it’s purely achieved by the novel use of diffusion models for code, that Stefano and two PhD student turned professor cofounders (Aditya and Volodymyr) invented years ago.

Maybe there’s a world where we DONT need billions in compute when you can run quality models cheaper ($0.25/M input, $1/M output tokens). The next step is to get them to be at frontier quality.

AI veterans Andrew Ng and senpai Andrej Karpathy are also investing alongside us at Menlo Ventures as well as NVIDIA, Microsoft, Databricks and Snowflake.

A small testament to just how incredible this team is the long list of incredible startups have already come out of Stefano’s lab: founders and key members of SSI, Pika, Luma, Together, Harmonic, Wispr, Liquid, Radical Numerics and many more!

Technical blog: https://lnkd.in/gtggzFFr
TechCrunch:
https://lnkd.in/gCzd6qT3

The biggest acquisition of 2025 and 5th biggest in Media of all time just happened with Netflix buying Warner Brothers!

For $82.7B, they'll acquire assets like HBO Max, Harry Potter, Game of Thrones and DC Comics Universe.

If you've not been following, here's everything that happened in the Anthropic x Pentagon x OpenAI saga.

This 41-page report on the business of tokens, both AI and crypto, is the best thing I’ve read this week.

It covers:
— How we will feed the machines
— How the machines will feed us
— What will power the machines

Link to the report (by Ribbit Capital): https://lnkd.in/gCV5PH6a

Wow, Gemini 3.1 Pro just crushed one of my test problems no other model could solve.

"Write a photorealistic 3D ocean simulation"

3D graphics is extremely hard. It has to encode 9-10 techniques based in physics. Being off by just 1 number can lead to bonkers results

The actual prompt was actually a bit more complex. It needed Gerstner Waves, Schlick’s Fresnel approximation, Blinn-Phong specular highlights, ACES Filmic Tone Mapping, Subsurface scattering, Fractional brownian motion, Wave grouping, Geometric splashing, Volumetric foam shading!

The one thing that didn't work were the Voronoi bubble trails!

Link: https://ocean.oneapp.dev/
Code: https://lnkd.in/gZjH2VKQ

I was wrong about Sarvam.

When I wrote about them a year ago, I felt like the direction to train small "indic" language models was wrong. But boy, have they turned it around. They have the best text-to-speech, speech-to text, and OCR models for Indic languages, and that's actually really valuable. The pricing is very reasonable. And the website is not only beautifully designed but dirt easy to use. They're filling a well needed gap in the ecosystem and doing things big labs will probably never focus on to the fullest extent (at least in the short term). I don't know anything about the business, but there's a lot to appreciate about what they've build technologically and I can't remember the last time I felt that way about software products coming out of India. Well done.

Few know that Vietnam is the world's most improved nation in the last 35 years.

Highest growing GDP per capita in the world from 1990-today, even more than China (only 2nd to Guyana, who discovered oil).

In that time, they went from ~4x less to ~2x more than India. Incredible.